Comparison between cephalometric measurements using digital manual and web-based artificial intelligence cephalometric tracing software

Abstract

Objective: The aim of this study was to compare the measurements performed with digital manual (DM) cephalometric analysis and automatic cephalometric analysis obtained from an online artificial intelligence (AI) platform, according to different sagittal skeletal malocclusions. Methods: Cephalometric radiographs of 105 randomly selected individuals (mean age: 17.25 ± 1.87 years) were included in this study. Dolphin Imaging software was used for DM cephalometric analysis, and the WebCeph platform was used for AI-based cephalometric analysis. In total, 10 linear and 12 angular measurements were evaluated. The paired t-test, one-way ANOVA test, and intraclass correlation coefficient tests were used to evaluate the differences between the two methods. The level of statistical significance was set at p< 0.05. Results: Except for SNB, NPog, U1.SN, U1.NA, L1-APog, I/I, and LLE parameters, all other parameters presented significant differences between the two methods (p< 0.05). While there was no difference (p> 0.05) in both SNA and SNB measurements between the two methods in the Class I malocclusion group, there was a difference between both methods in the Class II malocclusion group. Meanwhile, only the SNA in the Class III malocclusion group was different (p< 0.05). The ANB angle differed significantly in all three malocclusion groups. For both methods, all parameters except CoA and CoGn were found to have good correlation. Conclusion: Although significant differences were detected in some measurements between the two cephalometric analysis methods, not all differences were clinically significant. The AI-based cephalometric analysis method needs to be developed for more specific malocclusions.

Application of Artificial Intelligence (AI) in a Cephalometric Analysis: A Narrative Review

“…Table 1 presents the summarised studies on the application of AI in cephalometric analysis. In total, 23 articles were included based on both AI algorithms designed by their authors for the purpose of a specific study [ 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 ] and web-based software available on search engines and mobile applications [ 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 ]. The studies focused on comparing the reliability of AI algorithms in localising cephalometric landmarks on lateral cephalometric radiographs with the manual tracing of these points; differences between various algorithms were also examined [ 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 ].…”

Inteligencia artificial en diagnóstico, pronóstico y planificación del tratamiento de alteraciones de la región cráneo-cérvico maxilofacial en ortodoncia. Revisión de la literatura

“…La mayoría de artículos revisados van enfocados al diagnóstico; así mismo, los diferentes tipos cefalometría lateral de cráneo usan ciertos puntos anatómicos los cuales son una herramienta importante para el diagnóstico y la planificación de un tratamiento ortodóntico, ayuda a predecir algunos patrones de crecimiento de manera individual. En la actualidad se están realizando diversos estudios que proveerán de un sistema de IA exacto y confiable en el reconocimiento automático de los puntos y en la realización del diagnóstico (11,12).…”

“…Al comparar al trazado cefalométrico manual con una cefalometría obtenida de una plataforma de IA en línea; los software utilizados fueron el "Dolphin Imaging cephalometric análisis (v. 11.5, California, USA) y el WebCeph (WEBCEPH™, Artificial Intelligence Orthodontic & Orthognathic Cloud Platform, South Korea, 2020)", como resultado, se observó que en mal oclusiones clase I las medidas SNA y SNB no tuvieron diferencias entre ambos métodos, en pacientes de clase II hubo diferencias ambas medidas, mientras que en maloclusiones clase III solo el SNA fue diferente, solo los parámetros de Co-A y Co-Gn tuvieron una buena correlación, la cefalometría basada en (IA) necesita desarrollar un método más específicos en diagnóstico de maloclusiones clase II y III. (12). En pacientes clase III Hong et al mediante CNN usando en "Retina Net" para la detención de las regiones de interés y "U-Net" para la predicción de los puntos, en pacientes sometidos a tratamientos de ortodoncia y cirugía ortognática de ambos maxilares se asignaron variables en las mediciones y se consideraron parámetros tales como: excelente (menor a 1 mm), bueno (entre 1 a 1,5mm), justo (entre 1.5 y 2 mm), aceptable (de 2 a 2.5 mm) y no aceptable (mayor a 2.5 mm),así mismo, se evaluaron 12 marcas craneales y 8 detalles, este software tiene la ventaja que podría ser usado para la identificación de puntos en las radiografías a pesar de la presencia de brackets, placas y tornillos quirúrgicos, retenedores fijos, genioplastias y cambios de remodelado óseo, sin embargo la exactitud en algunos puntos no es lo suficientemente confiable para realizar un diagnóstico y planificar un tratamiento, en las marcas dentales mxc1 la corona del incisivo central maxilar los valores fueron 0.44mm y 97.8%, mxd6 contacto distal del primer molar mandibular fue de 1.43mm y 64.1%.…”

Influence of growth structures and fixed appliances on automated cephalometric landmark recognition with a customized convolutional neural network

“…As automated cephalometric software platforms are now available from different companies (e.g. OneCeph, Hyderabad, India; CellmatIQ, Hamburg, Germany; WebCeph, Republic of Korea; AudaxCeph, Ljubljana, Slovenia) more recent studies have focused on evaluating their accuracy [ 15 , 29 – 33 ]. While the benefits of artificial intelligence in recognizing cephalometric landmarks have been acknowledged [ 34 , 35 ], the need for further research regarding its accuracy in different clinical settings was recognized [ 36 – 38 ].…”

“…While the benefits of artificial intelligence in recognizing cephalometric landmarks have been acknowledged [ 34 , 35 ], the need for further research regarding its accuracy in different clinical settings was recognized [ 36 – 38 ]. Previous studies tested the frameworks only on radiographs of patients with permanent dentition [ 24 , 30 , 33 ] or did not mention these characteristic of the datasets at all [ 15 – 17 , 25 , 26 ]. Despite the promising potential of automatic landmark recognition, conclusions and research regarding some clinical aspects are still lacking.…”